How Resilient Are Kolmogorov–Arnold Networks in Classification Tasks? A Robustness Investigation DOI Creative Commons
Ahmed Dawod Mohammed Ibrahum, Zhengyu Shang, Jang‐Eui Hong

et al.

Applied Sciences, Journal Year: 2024, Volume and Issue: 14(22), P. 10173 - 10173

Published: Nov. 6, 2024

Kolmogorov–Arnold Networks (KANs) are a novel class of neural network architectures based on the representation theorem, which has demonstrated potential advantages in accuracy and interpretability over Multilayer Perceptron (MLP) models. This paper comprehensively evaluates robustness various KAN architectures—including KAN, KAN-Mixer, KANConv_KAN, KANConv_MLP—against adversarial attacks, constitute critical aspect that been underexplored current research. We compare these models with MLP-based such as MLP, MLP-Mixer, ConvNet_MLP across three traffic sign classification datasets: GTSRB, BTSD, CTSD. The were subjected to attacks (FGSM, PGD, CW, BIM) varying perturbation levels trained under different strategies, including standard training, Randomized Smoothing. Our experimental results demonstrate KAN-based models, particularly exhibit superior compared their MLP counterparts. Specifically, KAN-Mixer consistently achieved lower Success Attack Rates (SARs) Degrees Change (DoCs) most attack types datasets while maintaining high clean data. For instance, FGSM ϵ=0.01, outperformed MLP-Mixer by higher SARs. Adversarial training Smoothing further enhanced t-SNE visualizations revealing more stable latent space representations perturbations. These findings underscore improve security reliability settings.

Language: Английский

Machine Learning Applications in Building Energy Systems: Review and Prospects DOI Creative Commons

D. Li,

Zhenzhen Qi,

Yiming Zhou

et al.

Buildings, Journal Year: 2025, Volume and Issue: 15(4), P. 648 - 648

Published: Feb. 19, 2025

Building energy systems (BESs) are essential for modern infrastructure but face significant challenges in equipment diagnosis, consumption prediction, and operational control. The complexity of BESs, coupled with the increasing integration renewable sources, presents difficulties fault detection, accurate forecasting, dynamic system optimisation. Traditional control strategies struggle low efficiency, slow response times, limited adaptability, making it difficult to ensure reliable operation optimal management. To address these issues, researchers have increasingly turned machine learning (ML) techniques, which offer promising solutions improving scheduling, real-time BESs. This review provides a comprehensive analysis ML techniques applied According results literature review, supervised methods, such as support vector machines random forest, demonstrate high classification accuracy detection require extensive labelled datasets. Unsupervised approaches, including principal component clustering algorithms, robust identification capabilities without data may complex nonlinear patterns. Deep particularly convolutional neural networks long short-term memory models, exhibit superior forecasting Reinforcement further enhances management by dynamically adjusting parameters maximise efficiency cost savings. Despite advancements, remain terms availability, computational costs, model interpretability. Future research should focus on hybrid integrating explainable AI enhancing adaptability evolving demands. also highlights transformative potential BESs outlines future directions sustainable intelligent building

Language: Английский

Citations

3
IntelliGrid AI: A Blockchain and Deep-Learning Framework for Optimized Home Energy Management with V2H and H2V Integration DOI Creative Commons
Sami Saeed Binyamin, Sami Ben Slama

AI, Journal Year: 2025, Volume and Issue: 6(2), P. 34 - 34

Published: Feb. 12, 2025

The integration of renewable energy sources and electric vehicles has become a focal point for industries academia due to its profound economic, environmental, technological implications. These developments require the development robust intelligent home management system (IHEMS) optimize utilization, enhance transaction security, ensure grid stability. For this reason, paper develops an IntelliGrid AI, advanced that integrates blockchain technology, deep learning (DL), dual-energy transmission capabilities—vehicle (V2H) vehicle (H2V). proposed approach can dynamically household flows, deploying real-time data adaptive algorithms balance demand supply. Blockchain technology ensures security integrity transactions while facilitating decentralized peer-to-peer (P2P) trading. core AI is Q-learning algorithm intelligently allocates resources. V2H enables power households during peak periods, reducing strain on grid. Conversely, H2V facilitates efficient charging cars hours, contributing stability utilization. Case studies conducted in Tunisia validate system’s performance, showing 20% reduction costs significant improvements efficiency. results highlight practical benefits integrating technologies into innovative frameworks.

Language: Английский

Citations

1
Experience Knowledge Decomposition – Data Generation: Enhanced multi-step short-term cooling load predictions in data centres with data shortage issues DOI
Lei Zhan, G. Li,

Chengliang Xu

et al.

Energy, Journal Year: 2025, Volume and Issue: unknown, P. 136476 - 136476

Published: May 1, 2025

Language: Английский

Citations

0
How Resilient Are Kolmogorov–Arnold Networks in Classification Tasks? A Robustness Investigation DOI Creative Commons
Ahmed Dawod Mohammed Ibrahum, Zhengyu Shang, Jang‐Eui Hong

et al.

Applied Sciences, Journal Year: 2024, Volume and Issue: 14(22), P. 10173 - 10173

Published: Nov. 6, 2024

Kolmogorov–Arnold Networks (KANs) are a novel class of neural network architectures based on the representation theorem, which has demonstrated potential advantages in accuracy and interpretability over Multilayer Perceptron (MLP) models. This paper comprehensively evaluates robustness various KAN architectures—including KAN, KAN-Mixer, KANConv_KAN, KANConv_MLP—against adversarial attacks, constitute critical aspect that been underexplored current research. We compare these models with MLP-based such as MLP, MLP-Mixer, ConvNet_MLP across three traffic sign classification datasets: GTSRB, BTSD, CTSD. The were subjected to attacks (FGSM, PGD, CW, BIM) varying perturbation levels trained under different strategies, including standard training, Randomized Smoothing. Our experimental results demonstrate KAN-based models, particularly exhibit superior compared their MLP counterparts. Specifically, KAN-Mixer consistently achieved lower Success Attack Rates (SARs) Degrees Change (DoCs) most attack types datasets while maintaining high clean data. For instance, FGSM ϵ=0.01, outperformed MLP-Mixer by higher SARs. Adversarial training Smoothing further enhanced t-SNE visualizations revealing more stable latent space representations perturbations. These findings underscore improve security reliability settings.

Language: Английский

Citations

2